← Research Lab / Systems
FIG. 3261 // FANTASMA RESEARCH NOTE

Byzantine Consensus in Low-Power Edge Meshes: Revisiting PBFT for the Real World

Oleh Yuki Tanaka
14 min baca
2026-06-20

Byzantine Consensus in Low-Power Edge Meshes

Distributed systems research mostly assumes cloud conditions: low latency, reliable networking, ample compute. The edge is a different world — intermittent connectivity, heterogeneous hardware, strict power budgets.

When we built Edge Lattice, we needed a consensus protocol that could handle Byzantine faults (malicious or corrupted nodes, not just crashes) on a mesh where individual nodes might be solar-powered with 30-minute duty cycles.

What PBFT Assumes (and What We Don't Have)

Practical Byzantine Fault Tolerance (Castro & Liskov, 1999) requires:

Reliable eventual message delivery — *we have packet loss rates up to 8%*
O(n²) message complexity — *40-node mesh = 1,600 messages per round*
Deterministic timers — *edge nodes drift ±200ms per hour without NTP*

None of these hold in our deployment environments.

Our Modifications

1. Sparse quorum selection: Instead of broadcasting to all n nodes, we use a gossip-based approach to select a rotating quorum of √n nodes per round. Message complexity drops from O(n²) to O(n·√n).

2. Probabilistic view-change: Rather than timeout-based view-change (which misfires on slow links), we use a gossip-estimated liveness score per leader node.

3. Commitment DAG: We replace the linear commit chain with a DAG structure that tolerates out-of-order arrival without protocol violations.

Results on a 40-Node Testbed

ProtocolConsensus latencyMax Byzantine nodesPower/node
Vanilla PBFT180msn/3
Ours2.1msn/30.8W

Note: our latency advantage comes primarily from the quorum reduction — we're comparing different protocol shapes, not claiming a direct algorithmic speedup over PBFT.

Publikasi Riset Terkait

Architecture 11 min

Attention Is Not Enough: Toward Structured World Models at the Edge

Transformer attention scales quadratically with sequence length. For always-on edge inference, this is a thermodynamic impossibility. We propose a hybrid SSM-attention architecture that achieves 94% of full-attention quality at 8% of the compute.

HCI 8 min

Calibration-Free Gaze Tracking at 60fps on Commodity Hardware

Classical gaze tracking requires a 90-second calibration ritual that most users abandon. We trained a universal gaze model on 4.2M synthetic eye-renders that generalizes to unseen users with 0.4° accuracy — no calibration, no personal data.